It entirely depends on what you mean by “better at cubing” and “AI”. I’ve listed a few questions and answers below.
- Is there an AI which possesses the critical and creative thinking skills to create a method which surpasses what humans have invented? No, not yet at least.
- Can technology, including AI, be used in order to aid in method and algorithm development? Absolutely.
- Are there computer programs which can generate better solutions than humans can find? Yes, that’s been the case for decades.
They’re based on exhaustive search, heuristics, and optimization algorithms rather than human-style reasoning. You can read about them
here.
- Can ChatGPT find a better solution than you can? Doubtful; it doesn’t really even have a solid model of what a Rubik’s cube is and how it physically operates, and it won’t gain those skills unless it’s trained specifically for that purpose. The biggest thing it’s missing is an internal state tracker—this is also why LLMs get very confused by chess.
- Can robots solve Rubik’s cubes faster than we can? For sure; the
record for 3x3 is 0.103 seconds.
- Can you train a neural network to be very good at solving a virtual Rubik’s cube? Yes, and it’d be a super fun coding project! I recommend reading up on
AlphaZero, which is a chess bot which was able to beat the strongest chess engine in the world and was trained with zero outside instruction (other than being given the rules of chess and the ability to play games against itself) purely via reinforcement learning.
What you’re saying here is broadly true, but there’s a lot of inaccuracies:
- Programs like batchsolver don’t use AI at all, instead relying on deterministic algorithms and exhaustive search (incredibly optimized brute force, essentially). See my above point about them for more info.
- A lot of people like to simplify LLMs by saying “they just guess the next word” or “they can’t come up with anything new, just recycle old things.” These are gross oversimplifications and sometimes outright wrong. I don’t feel at all qualified to speak on the technical workings of LLMs, so instead I’ll link
this video by 3blue1brown and recommend you watch his other videos on the subject. Keep in mind when learning about AI that the field is changing rapidly and what was true 6 months ago could be remarkably outdated today. However, to the best of my knowledge, that video is a good basic starting point for getting an idea of how these models work.
- While I’m of the personal belief that there is a unique human element which cannot be fully replicated by machines, AI models are demonstrably capable of creating new ideas. Whether or not this is “actual thinking” and “true creativity” is a philosophical question (and quite a loaded one, at that). Regardless,
this is a great example of AI’s ability to be novel. Note the article is nearly four years old, and technology has advanced considerably even in that timeframe.
P.S: in researching this post (in an effort to not accidentally misinform), I came across
DeepCubeAI (based on the earlier
DeepCube[A] work), which is an utterly fascinating testament to how powerful heuristics and reinforcement learning can be. DeepCubeA was given a model of a Rubik’s cube and learned to solve it completely from scratch, and DeepCubeAI went even further and learned both how the cube works and how to solve it purely from visual observations. I wasn’t aware of these when I wrote the earlier point about creating a specialized AI to solve cubes, but still wanted to mention them. I recommend reading the linked papers if you’re interested in the relevant topics.
